Gå til indhold
Use casesLærOm mig
cleverest
Library

Nate Jones' prompting framework

Nate Jones' prompting framework

A job aid for prompting frontier models well. Source: Nate B Jones (AI News & Strategy Daily), "ChatGPT-5 prompting" manual and video (https://www.youtube.com/watch?v=hvTGYMq3pfg). Built around GPT-5, but the principles carry to any high-powered model (Claude included). Cleverest's contribution is packaging it as a job aid.

The one idea

A frontier model is a speedboat with a big rudder. Huge power, wants to go fast, needs hard steering. Feed it a vague one-liner and it fills the gaps by making things up — confident, detailed, useless.

The fix isn't writing more by hand. It's structure. And the practical shortcut to structure is a meta-prompt: a wrapper that turns your messy request into a proper brief, then runs it. Think of it as power steering — you write the way you naturally write, the wrapper does the steering.

Use it / skip it

Use the full treatment for real work with a deliverable — a meeting prep, a proposal, an analysis, a draft, anything where "done" has a shape and being wrong costs you.

Skip it for simple factual lookups, open-ended exploration, or emotional or reflective conversations. Casual prompting is fine there. (For emotional nuance, Claude beats ChatGPT anyway.)

The 7-component prompt checklist

Cover these and the model stops guessing. Not every prompt needs all seven, but the more it matters, the more you fill in.

  1. Role — name the expertise you need ("act as a B2B marketing strategist"). Not theater — it aims the model at the right knowledge.
  2. Objective — the mission. What's the goal? The model needs something to do.
  3. Process — the steps. "First do X, then Y, then Z." Give it a method, not just a target.
  4. Format — exact output shape. Email? Table? One-page memo? Bullet list? Say so.
  5. Boundaries — the anti-goals. What not to do, what to leave out, what to avoid. ("Don't invent statistics.")
  6. Uncertainty handling — what to do when it's stuck or data is thin. "If you're unsure, ask me before assuming." Rank goals if they conflict: "Primary is X; if X and Y clash, choose X."
  7. Validation — a way to check its own work. "List the assumptions you made" or "flag anything you couldn't verify."

The 7 principles behind it

Why these models misbehave, and the lever for each:

  1. Structure drives the answer — your headers and bullets shape how the model routes itself internally. Clear structure leads to better answers.
  2. Contradictions cost you (the "precision tax") — "be thorough but brief" makes the model burn time fighting itself. State a primary goal and a tiebreaker.
  3. Depth ≠ length — how hard it thinks and how long it writes are two separate dials. You can ask for deep thinking in a short answer. Specify both.
  4. It's literal — it will attempt anything, even what it shouldn't. Tell it explicitly where the edges are and what to do at them.
  5. Be opinionated about tools — it's all-in or all-out on web search and similar tools. Tell it when and how: "search first, then analyze."
  6. Its memory is an illusion — it re-reads everything each turn and over-weights your last message. In long chats, restate key instructions. (See the flag trick below.)
  7. Structure beats brute force — a clear method and shape get you further than trying to force "thinking mode."

The flag trick (catching memory loss)

Add this line to your opening prompt:

"If you've read and will follow these instructions, end every response with the word flag."

When flag disappears from the replies, the model has dropped your original setup. Time to restate it. This lets you see the moment it forgets instead of guessing.

The copy-paste meta-prompt

Paste this, then add your real request at the end. It expands your vague ask into a brief and executes it.

Transform my request into a structured brief, then execute it.

First, interpret what I'm actually asking for:
- What type of output would help me?
- What expertise is relevant?
- What format is useful?
- What level of detail?
State your assumptions so I can correct them.

Then restructure and execute with:
- A specific role (infer the right expertise)
- A specific objective (make my vague request concrete)
- An approach (pick the method that fits)
- An output (deliver it, leaving blanks where you'd otherwise guess — don't invent facts)

Finally, ask me the 2–3 questions that would most improve the result.

My request: [drop your messy request here]

The blanks-not-fabrication instruction is the key move — it stops the model inventing numbers and "facts" to look complete.

The bigger map

Jones nests prompting inside four disciplines, and argues most people only practice the first:

  1. Prompt craft — wording a single request well (this job aid).
  2. Context engineering — designing what the model can see and pull from.
  3. Intent engineering — being clear on the actual goal behind the ask.
  4. Specification engineering — defining "done" precisely enough to verify.

"The gap is 10x" — i.e. people who do all four get an order of magnitude more out of the same model.

Sources

Related pages